Introducing the SpaceNet Off-Nadir Imagery Dataset
Note: SpaceNet is a collaborative effort between CosmiQ Works, DigitalGlobe and Radiant Solutions hosted on Amazon Web Services as a public dataset. (To learn more visit https://spacenetchallenge.github.io/ ).
SpaceNet is proud to release the Off-Nadir Imagery dataset. This dataset was specifically built to explore advanced algorithms capabilities to process high off-nadir imagery. The dataset includes 27 WorldView 2 Satellite images from 7 degrees to 54 degrees off-nadir all captured within 5 minutes of each other. The dataset covers over 665 square kilometers of downtown Atlanta and ~126,747 buildings footprints labeled from a nadir image. It is now available for download — for instructions, see the SpaceNet Off-Nadir Dataset page
Why is this important? The ability to use higher off-nadir imagery increase’s the access footprint of a satellite which increase’s the ability for a satellite to revisit an area quickly. This enables a constellation of satellites to be more responsive to real-time events such as earthquakes, hurricanes or other natural disaster scenarios. For more information about the use of off-nadir imagery in disaster scenarios, read a great blog post from Maxar C.T.O Walter Scott which discusses some of the real-world use cases for off-nadir imagery.
The SpaceNet Off-Nadir Dataset
The SpaceNet Off-Nadir dataset is built off of a unique collect by WorldView-2 where it collected 27 images of downtown Atlanta collected in one pass on December 22, 2009. This generated off-nadir angles from 7.8 degrees to 54.0 degrees off-nadir. This summer, building footprints were then digitized by a Radiant Solutions team from the 7.8 degree off-nadir image to represent ground truth with as little distortion as possible.
All buildings were then labeled to the same specifications as in previous SpaceNet datasets and placed through a rigorous QA/QC process. The dataset has then been tiled into 450m x 450m (900x900 pixel) tiles for the challenge. All tiles are bi-cubicly resampled to have the same number of pixels in each tile despite having a courser native resolution due to high off-nadir angles. To learn more about the dataset visit the SpaceNet Off-Nadir Dataset page. It is released under a Creative Commons Attribution-ShareAlike 4.0 International License.
Looking at Off-Nadir Imagery
Images from farther off-nadir usually present with two types of distortion.
- Building tilt: The roofline is not inline with the building footprint and consequently the side of the building comes into view.
- Resolution degradation: As the satellite looks at the earth at more of an angle, the pixels get distorted and therefore resolution is degraded.
In the images to the left, the center building is ~46m long and ~24m wide with a parking lot situated to the left. Through the progression, you can see that as the angle increases, more of the side of the building is shown and the effect of building tilt increases.
The parking lot to the left of the building demonstrates how the resolution decreases as the off-nadir angle increases.
Both of these distortions will introduce problems when trying to develop machine learning algorithms to work at farther and farther off-nadir angles.
These issues will be tested and explored in the SpaceNet Off-Nadir Building Detection Challenge
The SpaceNet Off-Nadir Building Detection Challenge.
The SpaceNet Off-Nadir Building Footprint Extraction Challenge will launch October 19th and is designed to encourage the development of techniques that can handle high off-nadir imagery. The Challenge asks competitors to convert satellite imagery into usable building footprint vectors for three classes of imagery, Nadir (0–25 degrees), Off-Nadir(25–40 degrees) and Very-Off Nadir (40–55 degrees). For this challenge, the SpaceNet Building Footprint Metric will be used from past competitions. The F1-Score from each off-nadir region will be averaged together for a final winning score.
We hope you participate in this exciting challenge! Start by registering at the SpaceNet Challenge Page and then download the dataset to compete in the SpaceNet Off-Nadir Building Detection Challenge.
In the next couple of weeks SpaceNet Partners will be releasing several blogs that discuss different aspects of off-nadir imagery and a baseline algorithm for use in the challenge.